4 ms·

First, it’s probably quite possible to train a network to filter on AI vs. human generated. There are many visible signs for the untrained eye and probably many

First, it’s probably quite possible to train a network to filter on AI vs. human generated. There are many visible signs for the untrained eye and probably many more that are not visible to humans but can be picked up from the correlation. Secondly, there is still a big human selection process going on. Only the most interesting and coherent images will find their way onto the public internet. In fact, if you can automatically detect that these images are AI, then they can serve as an additional training signal to help teach the AI which of its outputs are most likely to delight the human.

webmaven 4y ago
> Secondly, there is still a big human selection process going on. Only the most interesting and coherent images will find their way onto the public internet. That's rather optimistic. Plenty of low quality images are generated and posted to the public Internet. What is true is that most of that generated junk doesn't have much of a shelf-life, which reduces the odds of it making it into the training data.